Terrain Recognizer
From camera input to a visible prediction.
Terrain Recognizer is a 2023 learning prototype connecting a Teachable Machine image model to camera input. The public implementation uses JavaScript and TensorFlow.js in the browser, loading an exported model and its metadata. The interface starts a camera stream, runs predictions and displays each class's score. The focus was learning how model loading, browser permissions and inference fit together, not establishing a benchmarked terrain classifier.
- MY ROLE
- Developer - learning prototype
- PROJECT STATUS
- Learning prototype
- BUILT WITH
- JavaScript / Teachable Machine / TensorFlow.js / HTML / CSS / MediaDevices API
A problem
worth exploring.
An exported image model is difficult to understand in isolation. Connecting it to real input and showing its predictions makes the behavior easier to observe and question.
A browser-based learning prototype that connects camera frames to visible classification scores, providing practical experience with model integration and inference without claiming measured accuracy.
How I approached it
- Load an exported Teachable Machine model and its class metadata in the browser.
- Request camera access and feed video frames into a recurring prediction loop.
- Display class names and prediction scores so the model's response is visible rather than hidden behind a single label.
What I took away
- Loading a model is only part of the work; camera permissions, asynchronous setup and feedback shape the experience.
- A displayed prediction score is not evidence of accuracy on unfamiliar inputs; evaluation needs its own work.